Decoding and encoding of neural signals for peripheral interfaces

(1)

TRABALHO DE GRADUAÇÃO

DECODING AND ENCODING OF NEURAL SIGNALS FOR PERIPHERAL INTERFACES

Por

Lucas de Levy Oliveira

Brasília, dezembro de 2015

(2)

UNIVERSIDADE DE BRASILIA Faculdade de Tecnologia

Curso de Graduação em Engenharia de Controle e Automação

TRABALHO DE GRADUAÇÃO

DECODING AND ENCODING OF NEURAL SIGNALS FOR PERIPHERAL INTERFACES

Por

Lucas de Levy Oliveira

Relatório submetido como requisito parcial de obtenção de grau de Engenheiro de Controle e Automação

Banca Examinadora

Prof. Antônio P. L. Bó, ENE/UnB Orientador

Profa. Mariana C. B. Matias, FGA/UnB Co-orientadora

Profa. Christine A. Coste, INRIA Sophia-Antipolis Examinadora externa

Brasília, dezembro de 2015

(3)

FICHA CATALOGRÁFICA OLIVEIRA, LUCAS DE LEVY

Decoding and encoding of neural signals for peripheral interfaces, [Distrito Federal]2015.

viii, 68p., 297 mm (FT/UnB, Engenheiro, Controle e Automação, 2015). Trabalho de Graduação – Uni- versidade de Brasília. Faculdade de Tecnologia.

1. Neuroengenharia 2. Engenharia Biomédica

3. Estimulação neural

I. Mecatrônica/FT/UnB II. Título (Série)

REFERÊNCIA BIBLIOGRÁFICA

OLIVEIRA, L. de L., (2015). Decoding and encoding of neural signals for peripheral interfaces. Tra- balho de Graduação em Engenharia de Controle e Automação, Publicação FT.TG-n^◦020, Faculdade de Tecnologia, Universidade de Brasília, Brasília, DF, 68.

CESSÃO DE DIREITOS

AUTOR: Lucas de Levy Oliveira

TÍTULO DO TRABALHO DE GRADUAÇÃO: Decoding and encoding of neural signals for peripheral interfaces.

GRAU: Engenheiro ANO: 2015

É concedida à Universidade de Brasília permissão para reproduzir cópias deste Trabalho de Gra- duação e para emprestar ou vender tais cópias somente para propósitos acadêmicos e científicos. O autor reserva outros direitos de publicação e nenhuma parte desse Trabalho de Graduação pode ser reproduzida sem autorização por escrito do autor.

Lucas de Levy Oliveira

SMPW Quadra 25 Conjunto 03 Lote 7 Casa B - Park Way.

71745-503 Brasília – DF – Brasil.

(4)

Dedicatória

Dedico este trabalho ao meu amado tio Gerson Oliveira Jr., que, apesar de todo seu apoio e carinho, infelizmente não pôde acompanhar o término deste projeto.

Dedico, ainda, às amizades que conquistei ao longo dos últimos anos dentro e fora da Uni- versidade, especialmente a minha irmã Ana Clara da Hora, por todo apoio e bons momen- tos que compartilhamos. Dedico a Anna Carolina Pinheiro, grande amiga e secretária do Departamento de Engenharia Elétrica, pelos conselhos, pelo suporte e, especialmente, pelos momentos de diversão, descontração e alegria abundantes. Dedico, enfim, ao professor Antô- nio Bó, pela grande amizade construída, por toda ajuda e paciência concedida e pelo grande aprendizado por ele proporcionado nos últimos anos.

Acima de tudo, dedico este trabalho aos meus pais, Rogério de Oliveira e Mônica Machado;

aos meus avós, Levy Machado, Maria Aparecida, Gerson Oliveira e Maria Amélia; ao meu irmão, Victor de Levy; aos meus tios, tias, primos e primas; e ao meu amor, Marina Moreira, por todo carinho, amor e suporte emocional recebidos por todos, que foram os mais funda- mentais para a realização deste trabalho.

Lucas de Levy Oliveira

(5)

Agradecimentos/Acknowledgements

Agradeço, inicialmente, a todos da minha família. Sem vocês, jamais teria tido condições de crescer academicamente e emocionalmente como cresci ao longo dos anos que precederam a realização deste trabalho. Agradeço aos meus pais pelos conselhos frequentes e por sempre confiarem em minhas deci- sões. Agradeço aos meus avós, tios e primos pelo apoio emocional e pelo orgulho que me fazem sentir pelo que faço. Agradeço ao meu irmão pelas boas risadas, pela preocupação e por sempre estar presente quando preciso. Agradeço, por fim, a minha companheira de aventuras, por sempre me ajudar a ser a melhor versão de mim mesmo e nunca me permitir duvidar de meu potencial.

Agradeço aos meus grandes amigos e colegas de curso e de laboratório por toda ajuda em projetos e em disciplinas, além da forte amizade construída – em especial a Ana Clara da Hora, Anna Carolina Pinheiro, André Vinícius, Bruno Noronha, De Hong Jung, Guilherme Anselmo, Lucas Fonseca, Marcela Carvalho, Matheus Portela, Miguel Paredes (que, inclusive, me ajudou com os procedimentos cirúrgicos das baratas) e Tiago Pimentel. Agradeço à Marienne Narváez e ao professor Emerson Fachin-Martins por toda a ajuda nos experimentos desenvolvidos no LARA. Agradeço, ainda, a Murilo Marinho pela amizade, pela ajuda em diversos aspectos e, inclusive, no desenvolvimento do código para o projeto desenvolvido no LARA.

Agradeço, também, ao corpo administrativo da Universidade de Brasília, por todo o suporte técnico concedido durante a graduação e pelo companheirismo em diversos momentos de dificuldade do curso, em especial aos atendentes da secretaria da engenharia elétrica Lúcio e Natália; porteiros dos Serviços Gerais 11 Elisângela e Sidney; e responsável pela limpeza da Faculdade de Tecnologia Maria José.

Agradeço ao corpo docente da Universidade de Brasília pelas lições de vida compartilhadas e conheci- mentos técnicos adquiridos, em especial a Antônio Bó.

I am also grateful for Backyard Brains’, NDI Digital’s and Ripple’s technical support for all the help during the development of this work. I would like to thank for the friendships developed during my time in Johns Hopkins University, specially my friends from the exchange program and my team from Infinite Biomedical Technologies, specially Martin Vilarino, Megan Hodgson, Rahul Kaliki and Nitish Thakor, my mentors and friends.

I am thankful for my friends at the University of Utah who helped me in several aspects during the development of this work, specially Mandi Peterson, Yiman Zhang, Kacey Gao, David Kluger and Katie Aiello. I would like to thank David Warren for the opportunity presented, for all the technical help, for the advices and for being a great host. I would like to thank Zack Kagan for his help during the experiments and all the advices for the project, besides his helpfullness in different other aspects. I would also like to thank David Page and Mitchel Frankel for all the help in providing advices and material for the research.

Lucas de Levy Oliveira

(6)

(7)

RESUMO

O campo da neuroengenharia cresce a cada dia com estudos promissores sobre suas diferentes áreas, sendo uma delas a estimulação elétrica funcional (FES). Para o avanço desses estudos, diferentes plataformas experimentais biológicas são utilizadas a fim de melhor entender o funcionamento do sistema nervoso e, eventualmente, poder intervir ativamente a fim de recuperar funções perdidas devido a patologias ou acidentes. Além disso, para o melhor estudo da área, diversos pesquisadores se utilizam de artrópodes como insetos. Estas cobaias são muito úteis, devido às suas similaridades ner- vosas com seres mais complexos e facilidade de manuseio e aquisição. Desta forma, além de serem utilizados em pesquisas que se utilizam de seus tamanhos reduzidos e complexidade biomecânica para realizar tarefas difíceis para a micro-robótica atual, os insetos também são utilizados para estudos introdutórios em neuroengenharia. Neste trabalho, duas plataformas experimentais relativas aos dois cenários descritos para estimulação em nervos são desenvolvidas e testadas – ambas serão aprofundadas brevemente a seguir, a começar pela última.

No Laboratório de Automação e Robótica (LARA da Universidade de Brasília (UnB), foi desenvol- vida uma plataforma experimental para implementação de algoritmos de controle de direção para baratas da espécieBlaberus giganteus. Essesetupcontou com o kit de desenvolvimento da Backyard Brains, o RoboRoach, afixado às costas da barata a ser experimentada para estimulação elétrica de nervos de suas antenas. A placa RoboRoach foi interfaceada com um computador pessoal (PC) por meio de protocolo Bluetooth (BT) 4.0, onde o algoritmo de orientação foi implementado e executado.

O PC foi também ligado a um sensor de captura de movimento da NDI Digital, o Polaris Spectra. Este sensor, por meio de emissão e recebimento de ondas infravermelhas, provia ao PC informações de posição do marcador passivo afixado à placa nos três eixos de deslocamento.

Desta forma, traçada uma referência retilínea, um algoritmo de orientação simples foi implementado de forma a estimular a antena da barata referente ao lado para a qual esta não deveria ir, fazendo uso de um comportamento evasivo gatilhado pelas antenas das baratas. Assim, a barata poderia ser orientada a seguir uma trajetória retilínea. Um controlador proporcional, ainda, foi implementado de forma que a amplitude de estimulação variasse de acordo com o distanceamento da barata em relação à referência desejada. Para os experimentos, foi utilizada inicialmente uma "barata virtual", referente à utilização de diodos emissores de luz (LEDs) dedebugpresentes na placa, que indicavam qual antena estaria sendo estimulada em determinado instante. Com isso, foi possível validar a efici- ência do controlador em si, isolando as variáveis referentes à estimulação da barata. Em seguida, os procedimentos cirúrgicos para experimentação com as baratas foram feitos e os experimentos foram executados.

Os resultados adquiridos dos experimentos mostraram que, apesar do algoritmo de controle se comportar bem em um cenário simulado (da "barata virtual"), houve impedimentos em relação à estimulação dos espécimes. Erros na forma como osetupexperimental foi desenvolvido e a estimu- lação foi feita foram encontrados e discutidos. Mesmo assim, resultados satisfatórios no controle de direção das baratas foram obtidos.

(8)

O trabalho foi, então, prosseguido no Centro de Neuroengenharia (CNE) da Universidade de Utah.

No novo setup experimental, foi desenvolvido um sistema em tempo real utilizado Matlab em um sistema operacional Windows 7. O sistema interfaceava com umhardwarede nome Grapevine, da Ripple LLC, que provia dados de eletromiografia adquiridos por uma de suas entradas e, além disso, possibilitava modulação de parâmetros de estimulação elétrica em uma de suas saídas. Desta forma, o sistema foi montado em diversas sessões experimentais: inicialmente, para validação do sistema em tempo real, se utilizando de sapos (de gêneroPtychadena) e, posteriormente, para desenvolvimento da plataforma experimental e implementação do algoritmo de controle, ratos (Sprague-Dawley). Am- bos os grupos de espécimes foram estimulados no nervo ciático utilizando um eletrodo do tipohook, de forma a desencadear uma estimulação de nervo completo (whole-nerve stimulation). Os sinais elétricos, assim, ativavam grupos musculares específicos e os sinais elétricos que resultavam em con- trações foram gravados e utilizados no sistema de controle. Osetupcontou, ainda, com a denervação de ramificações do nervo ciático, de forma a isolar o tipo de movimento causado pela estimulação, e com a consequente utilização de um sensor de força analógico para posterior comparaçãooffline com os sinais elétricos lidos do músculo (evoked electromyography– eEMG).

O algoritmo de controle proposto se utilizava de uma curva de recrutamento gerada por meio de variações de amplitude dos sinais de estimulação e o valor máximo absoluto (MAV) das respectivas respostas musculares. Esse procedimento era feito anteriormente à execução do controle, uma vez que este gerava a curva utilizada durante a execução do algoritmo de controle. Este consistia de lei- turas a cada 20m s (verificados em suas consistências posteriormente) dos sinais de eEMG, cálculo do respectivo MAV e utilização deste na função de controle. Anteriormente a esta, o valor atual de referência era computado, de forma a respeitar a forma trapezoidal desejada, de valores mínimo e máximo entre zero e 50% do valor de saturação da curva de recrutamento. O controle proporcional integral era, então, computado com base no erro entre valor esperado e referência e a variável de saída somada à referência desejada. Uma vez com o valor normalizado final, este era utilizado na função inversa da curva de recrutamento e o valor de amplitude a ser utilizado para estimulação era computado.

Foram adquiridos resultados referentes à consistência do sistema em tempo real e, ainda, à efici- ência do controle implementado. Apesar da discussão sobre consistência do período de amostragem se mostrar curta, uma vez que os 20m sesperados foram respeitados com uma pequena faixa de erro, muitos pontos em relação aos resultados obtidos do controle são discutidos. Qualidade da curva de recrutamento, efeitos da exposição do nervo ao ambiente, qualidade do controle utilizando diferentes ganhos proporcionais e integrais, relação entre eEMG lido e força resultante, estratégias de filtragem implementadas, fadiga muscular observada e erros de implementação do controle são discutidos. Por fim, conclui-se que, apesar do sistema experimental (hardwareesoftware) e do algoritmo de controle poderem ser melhorados no futuro, houve sucesso no desenvolvimento dosetupe na implementação do algoritmo de controle.

(9)

Palavras Chave: neuroengenharia, estimulação neural, eletromiografia evocada, controle PI

ABSTRACT

This work consists on the development of two experimental setups for the implementation of control algorithms in animals. The first one, developed at the Automation and Control Laboratory at the Uni- versity of Brasilia, consists on the electrical stimulation of antennal nerves of cockroaches (Blaberus giganteus) for direction control based on a retilinear reference. The position of the cockroach was measured by a motion capture system, which communicated with a personal computer (PC) running the control algorithm. The PC communicated with the stimulation board fixed to the cockroach back through Bluetooth 4.0 protocol. The control algorithm implemented was a simple line-follower which steer the cockroach by stimulating the antenna contrary to the desired direction with an amplitude proportional to the error between measured and reference positions. The second experimental setup took place at the Center for Neural Engineering at the University of Utah, where experiments using grass frogs and Sprague-Dawley rats were conducted. For both species, a hook electrode was used for whole-nerve stimulation of the sciatic nerve and wire electrodes were inserted into gastrocnemius muscles in order to record evoked electromyography (eEMG) singals. A proportional integral control algorithm was implemented in order to control muscle activation based on maximum absolute value (MAV) of the eEMG measured and recruitment curves found beforehand. Several aspects were discussed for both experimental setups regarding efficiency of the controllers and problems found in development and implementation. Finally, it is concluded that, even though there are several aspects to be explored in relation to improvements in both projects, the results found were satisfactory for initial trials.

Keywords: neural engineering, neural stimulation, evoked electromyography, proportional integral control

(10)

SUMÁRIO

1 INTRODUCTION. . . . 1

1.1 CONTEXT... 1

1.1.1 NEURALENGINEERING ANDINSECTS... 1

1.1.2 NEURALENGINEERING ANDFUNCTIONALELECTRICALSTIMULATION... 2

1.2 GOALS... 2

2 LITERATUREREVIEW. . . . 3

2.1 NEUROBIOLOGICAL BASIS OF MOVEMENT... 3

2.1.1 INSIDENEURONS... 3

2.1.2 BEYONDNEURONS... 5

2.2 INTERFACING WITHNEURONS... 8

2.2.1 LISTENING TO THENERVOUSSYSTEM... 10

2.2.2 TALKING TO THENERVOUSSYSTEM... 13

2.3 CONTROL OFARTHROPODSNAVIGATION... 17

2.3.1 PRINCIPLES OFNEUROPHYSIOLOGY ININSECTS... 18

2.3.2 INSECTSARTIFICIALNAVIGATIONUSINGNEURALINTERFACES... 19

2.4 FES CONTROLUSINGEVOKEDEMG ... 20

3 INSECTDIRECTIONCONTROLUSINGNEURALINTERFACES. . . 23

3.1 METHOD... 23

3.1.1 ROBOROACHDEVELOPMENTKIT... 24

3.1.2 POLARISSPECTRA ... 24

3.1.3 DEVELOPING OUR OWNAPI ... 26

3.1.4 COCKROACHPREPARATIONS... 27

3.1.5 THEEXPERIMENTALSETUP... 28

3.1.6 SIMULATIONS ANDEXPERIMENTS... 29

3.2 RESULTS... 31

3.3 DISCUSSION... 33

4 FES CONTROLUSINGEVOKEDEMG . . . 35

4.1 METHOD... 35

4.1.1 GRAPEVINE... 36

4.1.2 XIPPMEX... 38

4.1.3 DEVELOPING THEREAL-TIMESYSTEM... 38

(11)

4.1.4 CONTROLALGORITHM... 39

4.1.5 FROG ANDRATPREPARATIONS... 41

4.1.6 EXPERIMENTS... 43

4.2 RESULTS... 44

4.3 DISCUSSION... 51

5 CONCLUSIONS. . . 55

5.1 FINALCONSIDERATIONS... 55

5.2 FUTUREWORK... 55

5.2.1 INSECTSNAVIGATIONCONTROLUSINGNEURALINTERFACES ... 55

5.2.2 FES CONTROLUSINGEVOKEDEMG ... 56

REFERÊNCIAS BIBLIOGRÁFICAS . . . 58

ANEXOS. . . 65

I DESCRIÇÃO DO CONTEÚDO DOCD. . . 66

II PROGRAMAS UTILIZADOS. . . 67

(12)

LISTA DE FIGURAS

2.1 The diagram of a common action potential related with channels open. ... 4

2.2 Different waves found in electromyography signals. ... 6

2.3 The different types of electrodes in the order they were cited... 9

2.4 Block diagram describing preprocessing stages before decoding implementation. ... 12

2.5 Electrical stimulation parameters. ... 14

2.6 Expected effect of generated recruitment curve. ... 16

2.7 Example of remotely operated navigation control in a cockroach. Adapted from Latifet al[1]... 18

2.8 Basic neurophysiology of insects... 19

2.9 Cockroach atop a trackball used for sensing. Adapted from Holzeret al[2]... 20

2.10 Evoked EMG waveforms when stimulated by CIT and CFT. Adapted from Binder-Macleod et al[3]. ... 21

3.1 Block diagram describing the methods utilized... 24

3.2 The volume within which positions are acquired. Source: NDI Digital. ... 25

3.3 Final result from electrode implantation and board fixation procedure. ... 28

3.4 Arrangement used for the experiments. Polaris Spectra is positioned facing the floor, where a line indicates the desired path, namely, the reference path... 29

3.5 Control algorithm block diagram. ... 30

3.6 Voltage output measured from board with 1 and 10 pulses. ... 32

3.7 Voltage output measured from board with pulse-width of 1m s, 10m sand 50m s. ... 32

3.8 Simulations using the virtual roboroach setup. ... 32

3.9 Experiments using the real roboroach setup. In blue, the desired path the cockroach should follow. In red, the actual path traveled... 33

4.1 Block diagram of the simplified experimental setup. ... 36

4.2 Grapevine’s NIP and front-ends. Source: Ripple... 37

4.3 Block diagram of the control system implemented... 40

4.4 Experimental setup used in frog experiments. ... 42

4.5 Experimental setup used in rat experiments. ... 43

4.6 Force sensor attached to rat foot. ... 44

4.7 Force sensor (to the left) fixed to rat finger. The ankle position is fixed in order to acquire better measurements. ... 44

(13)

4.8 Loop periods during oscilloscope recording and stimulation. UDP communication was

turned on during the execution. ... 45

4.9 Loop periods during certain experiment on frog. UDP communication was turned on during the execution. ... 45

4.10 Data recorded during initial script running on second frog experiment... 46

4.11 Loop periods during certain experiment on rat. UDP communication was turned on during the execution. ... 46

4.12 Representation of evoked EMG recorded for single-reference (above) and differential electrodes (below)... 47

4.13 Evoked EMG in rat zoomed in; long experiment for fatigue perceiving. ... 47

4.14 Recruitment curves before and after electrode repositioning. ... 48

4.15 Respective control output for first recruitment curve in Figure 4.14. ... 48

4.16 Recruitment curve found in second leg. ... 49

4.17 Control output before and after moving window filtering. ... 49

4.18 Evoked EMG during control execution and respective force measured by transducer. ... 50

4.19 Control outputs for differentk_p gain values. ... 50

4.20 Control output recorded in last experiments... 51

4.21 Control output recorded in very last experiment. ... 51

(14)

LISTA DE TABELAS

3.1 Values used during experiments with cockroaches... 31

(15)

Chapter 1

Introduction

Biotechnology is not inherently good or bad; it is simply a set of techniques, and we have choices about how we employ them. If we use our scientific superpowers wisely, we can make life better for all living beings—for species that walk and those that fly, slither, scurry, and swim; for the creatures that live in scientific labs and those who run them. So it’s time to embrace our role as the dominant force in shaping the planet’s future, time to discover what it truly means to be stewards. Then we can all evolve together. — Emily Anthes, Frankenstein’s Cat[4]

1.1 Context

As both medicine and rehabilitation engineering advance their knowledge and techniques regarding the amelioration of neurological dysfunctions, no definite cure for certain pathologies and impair- ments, such as paralysis and amputation, have been developed yet[5, Foreword]. In this sense, a new field arises namedneural engineering, defined as "an emerging interdisciplinary research area that brings to bear neuroscience and engineering methods to analyze neurological function as well as to design solutions to problems associated with neurological limitations and dysfunction"[6].

1.1.1 Neural Engineering and Insects

Several approaches have been used in order to better understand the nervous system as a whole, from experiments with humans and other mammals, to the usage of less complex animals in experiments. Some of these researches have found in the study of insects a good approach in both explor- ing nervous system and development of important applications. Hence, the study of insects nervous system has been tied to two main factors: the development of instrumentation and algorithms implementation for interfacing with nervous systems and the usage of these strategies allied with insects natural minimized size and complex biomechanical responses needed for specific tasks.

Regarding the last aspects, the alliance between insects capabilities and recent technology has

(16)

resulted in several interesting applications, such as the usage of controlled-flight beetles that could be used as micro air vehicles (MAVs) with better miniaturization, payload and performance[7]; moth swarming for search and rescue in inhospitable areas[8]; and terrestrial insect navigation controlling for disaster sites exploration and search for possible victims[9]. Thus, it is understood nowadays the advantages of these researches, as they allow us to understand aspects of neural engineering appliable to other animals and, also, provide real-world applications.

1.1.2 Neural Engineering and Functional Electrical Stimulation

Neural engineering has seen three main fields emerging throughout the last years: stereotactic and functional neurosurgery,neuromodulationandfunctional electrical stimulation(FES)[10, p. 1- 1]. Though the first two fields bear their great significance, we will focus mainly in the latter for this work.

FES is a method consisting of applying low intensity electrical currents in order to restore or im- prove specific functions[11, Introduction]. Most of these functions are related to muscle activity control, such as bowel and bladder control, even though some relate to visual and auditory prostheses[10, p. 1-2]. In this work, the control of skeletical muscles will be the main focus. As just noted, FES does not aim at reversing paralysis once the system is turned off; however, in some applications, the user may receive benefits regarding the recall of voluntary muscle functions. Thus, among several other uses, it has been used in patients in order to maintain active muscle activity during physiotherapy sessions.

1.2 Goals

The goal of this work is to explore the neural engineering realm by working with different animals while developing different applications. We propose a navigation control for cockroaches using neural engineering techniques and a motion-capture sensor in order to trace paths on which the insect should thread. Also, a FES system is developed in order to control muscle activation in small animals, such as frogs and rats, by using whole-nerve stimulation and muscle electrical signals acquisition.

(17)

Chapter 2

Literature Review

There is only one nature – the division into science and engineering is a human imposition, not a nat- ural one. Indeed, the division is a human failure;

it reflects our limited capacity to comprehend the whole. — Bill Wulf

2.1 Neurobiological basis of movement

The union between engineering and neuroscience, although having shown to be benefitial for humankind, requires for one to understand basic principles of the other. Thus, while preoccupied with a world of equations and algorithms, engineers must also plunge into a more biology-centered domain and comprehend principles related to neurons and their properties. In order to act upon the system, the engineer must understand it – as the next sections should be helpful with.

2.1.1 Inside Neurons

One cannot understand the signals that are read in nerves or the brain without having a primary discerning of what a neuron is or does. A neuron – or brain cell – differentiates itself from a usual cell by being excitable or, more specifically, by specializing in the propagation of electrical current[12, p. 193- 194]. The opening and closing of certain electrically charged elements channel an electrical stream along the cell. The dendritesserve as the neuron inputs and drive the current passing through the soma– the cell body – into theaxons, which are considered the outputs. However, this current is only fired through the soma if the inputs exceed a certain threshold. This flow of charge is usually called anaction potential– commonly used when referring to the waveform – orspike– which represents the action potential in a wider time-scale.

Due to the characteristics of the ionic channels and their activation timings, action potentials have very specific waveforms depending on several factors, from the cell’s own properties to the position of the neuron related to the recording instrument[13]. The main qualities of spikes areamplitude, duration andfiring rate. All of these characteristics are going to be primal when dealing with the stimulation parameters. For now, it is necessary to understand them carefully.

(18)

Amplitude refers to the highest voltage value the action potential achieves within its duration. This duration is named pulse-width, which contemplates, in time, all voltage values from the beginning to the end of the wave. Spikes, however, do not always present themselves in a single firing, but in aspike train, which is a burst of several spikes evoked in sequence. The period at which they are discharged is the inverse of the spike frequency.

Another essential attribute of action potentials is their biphasic polarization. As the typical sequence of events takes place, when a conducting region of the neuron is electrically excited, specific channels activate, generating an inward flow of positive charges. Thisdepolarizesthe cell, which is originally around negative 70m V (theresting potential), activating another set of channels responsi- ble for the outflow of positive charges. This creates a steep decrease in voltage (repolarization), reach- ing a negative value even lower than the resting potential[14]–in other words, ahyperpolarization– resulting, thus, in the biphasic shape. After all channels are closed, there is an specific delay within which the cell cannot be stimulated, calledrefractory period.

depolar ization

repolar ization

hiperpolarization mV

membranevoltage

time

resting potential Na+inflow starts

K+outflow stops

resting potential

-70 0

K+outflow starts Na+inflow stops

ms

Figure 2.1: The diagram of a common action potential related with channels open.

The action potential then travels along the cell until the end of the axon, where it excites calcium channels, guidingsynaptic vesiclesto the membrane. These vesicles release neurotransmitters that will excite the next cell in thesynaptic cleft[15, p. 160-166]. This pulse travels along the nerves, from the brain to the nerves andvice versa.

(19)

Therefore, when dealing with motor and sensory interfaces, there are two well-known pathways followed by the signals: from the motor cortexto themotor units and fromskin sensorsto theso- matossensory cortex.

2.1.2 Beyond Neurons

In order to understand the usual pathways studied by neuroengineers, it is important to acquire knowledge about the routes, structures and the signals that interest the engineering realm. Neuronal cells are, hence, zoomed out to nerves, lobes and cortices.

2.1.2.1 The Brain Moves

Starting from the first path aforementioned, most signals that are able to actually move body parts originate primarly from thepremotor cortex. Within it, the brain designs how movement will be exe- cuted, although there are evidences that it is also involved in monitoring of accomplishment and performance[16, p. 255]. Signals are delivered to the motor cortex, which translates movements planned in premotor cortex to a more low-level and specialized context. These signals travel down the brain, passing through thecorticospinal tractand, finally, thespinal cord, where they branch out to different peripheral nerves[16, p. 686-690]. The spinal cord plays, thus, the role of a nerval hub–besides being involved in different types of autonomous movements by providing simple circuitries involving both efferentandafferentpathways[16, p. 558]. Inside these nerves, efferent fibers carry motor-related signals – as opposed to the afferent pathways, which carry sensoring information to the higher-up levels [16, p. 675-678].

Each of these efferent motor fibers, then, innervates different muscle fibers. Every single one of these groups of connections are calledmotor units. At this point, it is interesting to point that, usually, fast-acting and accurately controlled muscles have a large nerve-fibers-to-motor-fibers ratio, whereas slow-acting muscles who do not require precise control have a smaller ratio[16, p. 81]. Once the signal arrives at the muscle,actinandmyosinfilaments contract in an intricate mechanism, which generates force and contracts the muscle. The resulting coupled system of tendons and bones generate motion as we know it; however, it will not be thoroughly detailed in this document.

As the neural signals, who originated far back in the premotor cortex, travel through the muscle fiber, a signal can be recorded for analysis — this is calledelectromyography(EMG)[17, p. 179- 180]. EMG signals can be divided in two branches: those which are part of a natural process, as being described so far, calledvolitional EMG, and those that are result of electrical stimulation, called evoked EMG(eEMG). Even though volitional EMG is described merely by voltage, evoked EMG has specific features which differ from the latter. For example, there are two main components that can be assessed when measuring eEMG. The first one, calledH-wave(or Hoffman reflex, H-reflex), travels within the same pathways asmuscle spindles(sensors located inside the muscles that translate length into electrical activity). H-waves are typically easily excited, meaning their threshold for electrical stimulation is much lower. The second main component of EMGs is theM-wave, which, opposingly, is excited by higher signals. This wave is usually more closely related to the muscle contracting be-

(20)

havior itself, being the one most commonly used in motor analysis. Apart from the other two waves aforementioned, there is a third type, called theF-wave, which is generated once the motor fiber is hyperexcited, producing afferent and efferent stimuli, the latter generating the M-wave. The former travels through the nerve back to the spinal cord, where it returns as the F-wave, contracting the muscle after the M-wave does so[18, p. 127].

afferent efferent

m-wave h-wave f-wave

Figure 2.2: Different waves found in electromyography signals.

So far, it has been described the behavior of the motor system when a single signal comes forth.

This signal travels through the pathways described and, in the end, generates a singlemuscle twitch, which is a quick contraction of the muscles, which return to the original length due to their passive stiffness. In everyday experiences, however, it is known that this is not the usual performance of mem- bers – as they are able to mantain the muscles contracted by long periods until they can no longer be held steady. In order to do so, the neural system has to fire the motor units quickly and repeatedly in a short timespan. As explained above, this form of activity is described as a spike train. The muscle, acting as a low-pass filter, holds the spike for a while once it is excited. Accordingly, in order to maintain the same contraction fashion for some time, the firing rate increases to frequencies close to 50H z, resulting in what is referred to astetanic contraction– ortetanus. Obviously, the comparison with everyday living falls short to explain everything, since the recruitment of different muscles within the samemuscle groupare not being considered. On the other hand, it is a good instrument to easen the comprehension of other factors, such as the muscle’s inability to hold contraction for a long time [5, p. 39].

After this prolongued muscle work, the muscle does not present itself as capable of achieving the same amount of contraction than before, in a situation regarded asmuscle fatigue. Physiologically, it is involved in several situations. Fatigue can occur, for example, by interruption of blood flow, resulting in a decrease of nutrients necessary for muscle contraction. Another example is in prolongued activation and depletion of nutrients which were not consumed by the body – which is a common cause in laboratory experiments[16, p. 82]. In this situation, yet, it is most probably explained by the necessity of engaging in an active muscle process namedrelaxation, which consumes, besides time,

(21)

ATP (adenosine triphosphate, the substance involved in muscle contraction itself and in general active processes around the body)[19, p. 683]. It is interesting to bring out how muscle fiber types can affect muscle fatigue, being divided by two main kinds: slow fibersandfast fibersmuscle fibers. The first one is characterised by the presence of small fibers and great amounts of capilarity. This results in two main aspects: a smaller generation of force and better resistance agains fatigue. In counterpoint, fast fibers are acknowledged as larger – hence, greater force output – and with lesser blood supply, which reveals to be less resistant agains fatigue[16, 5, p. 80-81, p. 43]. Other nomenclatures when refering to these fibers are, respectively,type Iandtype IIorredandwhitemuscle fibers – the former pair refering to the presence or absence of hemoglobin, which acts as a source of oxygen and, thus, fatigue resistance[20, p. 17].

As fatigue, there is another muscle time-varying characteristic, themuscle potentiation, which is caused by increased sensitivity of calcium ionic channels after previous stimulations. This results in a substantial increase of muscle output force and stiffness[21]. All these nonlinear characteristics are truthfully important when designing control interfaces, which will be discussed more deeply further on.

2.1.2.2 The Brain Feels

As there are specific neural pathways for carrying information such as motion, there are – just as important – sensory pathwaysnecessary for our everyday activities, such as temperature, pressure, vibration and even pain perception. These specific variables are usually mediated byexteroceptors, that is, receptors or skin sensors that perceive external stimuli. The other main group of receptors is namedproprioceptors, relating to receptors that provide information about the human’s own body, such as muscle length, speed, contraction or even joint angle. This type of information is essencial for the brain to be able to finely control motion, serving as closed-loop feedback. This distinction, some- times, may fall into gray areas, as exteroceptors may all the same contribute to motion, for example, in high-temperature reflex arcs[19, p. 147].

There are other types of receptors in the body that should be mentioned, such as visual receptors in eyes and auditory receptors in ears. The receptors mentioned before, nevertheless, are known assomatic receptors. The usual pathway for their signals is to travel in afferent neurons through peripheral nerves into the spinal cord (where some of the sensors are processed in basic circuitries) and towards different areas of the nervous system, including: the medulla, pons, mesencephalon, cere- bellum, thalamus and, finally, the somatosensory cortex[16, p. 555].

Despite of the long list of sensory receptors present in vertebrates, this reality is a tad different when dealing with insects. While keeping the terminology of exteroceptors and proprioceptors, some sensors vary according to arthropod’s inner characteristics, such as the presence of an exoskeleton.

Taking the cockroach’s antenna (which is a very interesting subject to study, as it will be noted later) as an example, most of the proprioceptors are actually located externally to the body and can act as exteroreceptors. In other words, while somemechanoreceptors, receptors that react to mechanical changes – such as thecampaniform sensillaor thearticulated sensory hair– but only act as exteroceptors[22, p. 78], others can also work as proprioceptors, e.g. hair plates[23]. One of the most im-

(22)

portant aspects of the cockroach’s antenna, in fact, is a characteristic noticeable in different insects:

active sensoring. Because of its innervated nature[24], these insect parts can move in different direc- tions indepedently and be used in different sorts of behaviors[25]. The number of different actions is surprisingly long (including anemotatic behavior, wall-following and electrostatic field detection), the one with most interest for this work being theevasive behavior, which takes place when a cockroach perceives a strange obstacle as a predator – resulting in a quick turn-and-run acting[26].

2.2 Interfacing with Neurons

A lot has been told about the physiology involved in afferent and efferent pathways, motor units, muscles and sensory receptors; therefore, it is time to discuss about how these signals can be recorded, manipulated, generated and used for tightening the interface between man and machine.

It is primarily important to understand the basic terminology used in this sort of interface.Neural recording, for example, is the name of the technique used for acquiring neural signals, whereasneural stimulationis how machines interface with the nervous system, generating waveforms similar to biological ones and exciting the nerves from an external end. As interfaces vary, these names will keep their main name, as in "EMG recording" or "muscle stimulation". Moving on, this physical interface between machine and human is called anelectrode, which comes in different sizes and shapes. Usu- ally, there is an explicit trade-off between how invasive the electrode is – how complex is the surgical proceeding to implant it or not – and how selective it can be – namely, how many different neurons are interfaced with a single channel[27].

There are a few types of electrodes that can be mentioned:

• superficial electrodes, which provide a less invasive and less selective interface (generating lots of crosstalk, i.e., when two different muscles are recorded by the same channel[28]), being placed usually on the skin of the user;

• hook electrodes, which are designed as two or more hooks (depending on the number of chan- nels) on which the nerve is "hanged", discarding the electrical noise caused by tissues and the skin-electrode interface that commonly poses as a problem for superficial electrodes users[29], besides having better fixation than the latter[30, p. 45];

• cuff electrodes, whose shape follows the nomenclature by covering the nerve, guaranteeing a better-yet-not-enough selectivity;

• Longitudinal Intrafascicular Electrodes(LIFE), which are single wires inserted into the nerve in a longitudinal fashion;

• TIME electrodes, thetransversal intrafascicular electrodes, which follow the same logic as the latter mentioned, only transversaly[27]; and

• Utah Electrode Array(UEA), the one that has become more and more famous by the day due to its increased selectivity, number of electrodes and reach within peripheral nerves[31].

(23)

invasiveness

selectivity surface electrodes

TIME electrodes

Utah electrode array

LIFE electrodes

cuff electrodes hook electrodes

Figure 2.3: The different types of electrodes in the order they were cited.

An important characteristic that accompanies electrode selection when stimulating iscurrent density. When delivering an specific amount of charge to the nerve, the gauge through which this current will run is of great importance. In case a high charge is forced upon a narrow-diameter interface, the effect might involve dispersion through heat, which can cause pain to the patient and even harm the tissue, which is not at all advisable. By the same token, if a superficial electrode of broad contact area is selected, for example, the current may be diffused, resulting in a less-than-ideal stimulation threshold[5, p. 209-210]. Another aspect worth mentioning is that both recording and stimulation usually happen in asingle differential configuration, that is, a channel is always used for either stimulating or recording, whereas at least one channel must be set as reference[32, p. 274]. The signal measured, then, will be the difference between the signals measured by each electrode. In case of EMG (or evensuperficial EMG– sEMG – where wavelengths may be larger than the interelectrode distance), this reference can be used with more channels, allowing a technique namedspatial filter- ing, that significantly decreases low frequency artifacts – which will be detailed further on[5, p. 400]. As discussion over the motor system preceded, this very same order will be followed as related techniques are explained, allowing the phisiology discoursed to serve as an example. When interfacing with the motor system, many applications involve recording signals from either the nerves or the muscles themselves, as it follows.

(24)

2.2.1 Listening to the Nervous System

Whether one talks about common interfaces as sEMG-interfaced commercial prosthetic hands, more rare practices as targeted muscle reinnervation[33, p. 300], or even more complex procedures as surgical placement of an electrode array into a nerve for controlling a robotic hand, one thing is certain: how signals are read and recorded must be thoroughly understood.

Once the desired electrode has been decided through research and the counter-balance of advantages and disadvantages, the electrode must be placed on the appropriate location. This involves studying the neurophisiology of the species at hand and physical particularities of the subject, as well as the most suited site for the current application. For this goal, all sorts of equipments are used, such as oscilloscopes, muscle stimulators and, as strange as it may sound, audioamplifiers[5, p. 159].

The next main step involvesamplifyingandfilteringthe signal. Since signals recorded from these types of interfaces usually wander from the microvolts to the millivolts ranges, these signals must be magnified in order to be processed by usual electronics. For example, in EMG amplification, a series ofoperational amplifiersmust be used in order to attain a good signal quality, usually represented by high input impedance, high common mode rejection ratio and low noise. This series can be further described as the association of voltage followers and instrumentation amplifiers[32, p. 275]. Once the signal is amplified by tens or hundreds of its value, it is valuable to apply usual filters in order to decrease overall noise. One of the filters that may be used, as usual, is anotch filterof either 50H z or 60H z as the cutoff frequency. However, in case of neural recording, the signal main frequencies lie between specific ranges, such as 150H z–15k H z, allowing a simpleband-pass filterto be used[34, p. 483]. In case of EMG, the same band-pass filter can be implemented, however changing its cutt-off frequencies to the range of 1H z–3k H z[34, p. 557].

While discussing signal preprocessing, it is worthy mentioning how different external stimuli may interfere with the recorded signal. The stimuli mentioned, in fact, does not need to be of electrical nature only; in other words, these disturbances may happen because of concurrent electrical stimulation or even movement-related aspects. The subject in matter isstimulation artifactsandmotion artifacts. The latter may be present, for example, during sEMG acquisitions, when fixation of the superficial electrodes is not optimal, resulting in distinctive disturbances in the waveform. For this specific case, it has been proposed as solution the light abrasion of the skin surface using a fine sand- paper[35]. However, for artifacts produced by simultaneous electrical stimulation, the approach may involve less physical elements. This will be discussed along with the processing of eEMG signals soon.

Once the signals read are properly amplified and mostly free of noise and artifacts, hopefully by hardware implementation approaches, it is time to convert them to digital data (suitably using an A/D converter) and actually understand their meaning. The two major signal types usually recorded, notwithstanding, have different processing procedures, which will be detailed followingly. Of course, the types being mentioned are neural signals and muscle signals (eEMG), which shall be discussed in order. Neural signals waveforms have been explained earlier in this report, thus, from this stage onward, the plan is to extract those action potential waveforms and start thedecodingprocess, which consists in translating the characteristics of the spikes (usually the firing rate) into information that can be used in applications, such as joint angle and skin pressure. The first software implementation

(25)

of the input signal is calledspike detecting, which consists in actually finding the times at which a spike occurred. In order for this to happen, different algorithms can be implemented; yet, a somewhat simplistic approach is usually chosen in spite of the others due to its computational efficiency and ease of implementation: thethreshold detection. Because of action potential’s natural high amplitude peak, a voltage threshold trigger can be used to detect spikes. Thus, once the most proper value for the dataset is set – which can be found either manually or automatically during an offline analysis – the great majority of spikes will be found by this algorithm, since they tend to maintain their characteristical waveform, provided that source location is kept constant[36, p. R57]. This path has been continued by others, who have developed a double-threshold algorithm, consisting of positive and negative thresholds (contemplating both anodic and cathodic polarizations), which may decrease the ratio of false detections, along with filtering implementations[37].

Having found the action potentials or, rather, being able to do so for the following ones, it is time to actually classify the waveforms extracted. This is important since electrode channels are not always capable of accurately pinpointing a single neuron (being addressed thus as anextracellular record- ing), which would provide the waveform that has been made familiar in this document, but rather multiple different signals and waveforms. The general shape of the action potential is known; however, depending on geometry of electrode and distance from recording point, the signals may vary in amplitude or pulse width[38]. Hence, sorting algorithms are implemented, which may rely on two aspects: the overall form of the spike (template matching) or specific properties of the signal that are used in order to group them in similar categories, or neuron sources (feature clustering).

The first group of algorithms work basically by using predetermined spike templates, found some- times by creating a "mean waveform" gathered from previous samples and, added to white noise, find the probability that each template is found in the dataset. The second group extracts features from the waveforms, such aspeak time,peak-to-peak voltageorsettling timeand clusters the signals based on these attributes[38]. Other approaches, still within the second group, rely on acquiring the eigen- vectors of the waveforms set covariance matrix. This is calledPrincipal Component Analysis(PCA), a method that finds the orthogonal basis vectors that follow the largest variation within the data, i.e., an algorithm that automatically selects the best features to use when classifying groups of signals[36].

PCA usually does well in matters of spike sorting; however, it may be not enough in specific cases, such as when the main variation of the signal is in time, not in amplitude. For these situations, algorithms which deal better with time-scaling have been tested and shown to work better, such as the usage of wavelet coefficients[39]. Once these features were extracted, different kinds of clustering algorithms can be used, from eyeballing the regions and setting thresholds to algorithms such asK-meansor even Superparamagnetic Clustering(SPC)[39].

(26)

amplifying (instrumentation

amplifiers)

spike detection (threshold

crossing) feature

extraction clustering

artifact removal bandpass

and notch filters

spike counting (firing rate calculation)

Figure 2.4: Block diagram describing preprocessing stages before decoding implementation.

The following step from having classified the action potential according to their origin is to finally acquire the firing rate of each neuron and try to correlate it with variables to which it is wanted to translate. In other words, anestimatorthat will interpret the firing rate (which can be done by counting the number of spikes within a moving window) and output information about, for example, how contracted is the muscle, whether there is pressure on the skin or not, or even what temperature the body is sensing in an specific place. Followingly, neural signals will be placed on hold, as muscle signals treatment are discussed.

The first thing that should be noted when understanding the procedure of treating evoked EMG signals is that, since it originates from muscles, it will most probably be elicited when a signal generates a contraction. For this to happen, a stimulating signal must travel through the nerve, as it has been discussed before. In other words, when an electrode reads the muscle signals, it may also read the electrical stimulus as well; thus, strategies must be implemented in order to treat this titled artifact. Two known approaches are the usage ofblanking windowsandfinite impulse response filters.

The former one consists on nulling the signal read for an specific period at which the stimulation is present, whereas the latter consists on applying a filter to the whole recorded signal, which can be treated previously by the first approach[40]. While these methods were implemented in software and may be more time costing, other methods have been implemented in hardware, such as thefast settle triggersfrom Grapevine[41, p. 32]

While it is not the case when dealing with evoked EMG, one may notice how discontinuous the volitional EMG signal shows to be; instead of seeming well-behaved as shown in Figure 2.2. In order to achieve the signal mentioned, another processing must take place. After the signal has been properly rectified via either software or hardware (preferably by the latter), alinear envelopeshould be applied to the EMG signal, which will, as the name suggests, interpolate the peaks of signal, creating a smoother curve. Despite of some studies being able to show how to implement this method in a man- ner that is fast and efficient enough to be used in realtime[42], other methods may still be used online, such as peak-to-peak amplitude and mean amplitude value. Once this waveform is achieved, the am-

(27)

plitude of the volitional EMG can be used for implementing control algorithms to activate prosthetics [42], for example.

The procedures aforementioned are the ones which are commonly implemented in order to use decoding and control algorithms. For the next part, methods used for encoding signals, i.e., providing action potentials that can be understood by the nervous system, will be explained.

2.2.2 Talking to the Nervous System

Now that the mysteries behind data acquisition from the nervous system have been made a bit more clear, it is time to understand how one can emulate neural signals into it. As it has been discussed before, action potentials have specific shapes that should be recreated in order to generate an ideal response. Whether this response will go to efferent pathways and stimulate the muscles or to afferent pathways and stimulate the brain in order to modulate specific sensations such as heat and vibration, orientations should be followed in order to obtain better responses. For this goal, a few properties of neural stimulationwill be discussed next.

One of the first properties of the signal are related to the waveform, such as the signal shape:

whether it is rectangular, sinusoidal, triangular, etc. It is shown that the shape must affect more directly in the rising time of the signal, which should be fast enough for the nerve membrane not to acommodate and open the channels. For this purpose, most commonly arectangular-shaped stim- ulusis prefered. Another feature is related to whether the signal isbipolar ormonopolar, meaning whether the signal changes polarization within its pulse-width (resulting in a bipolar pulse) or not (monopolar pulse). The former is usually chosen due to better comfort and tissue integrity issues, besides preserving the electrode capabilities due to the balancing of charges[5, p. 209]. Another feature that may be added to the waveform is aninterphase delay, which is represented as an interval between the primary and secondary pulses. This delay may result in a decrease of the threshold separation, meaning a decrease in the slope of the recruitment curve[43], which will be better detailed later on.

(28)

mA

ms

µs

mA

period=1/frequency

duration

amplitude

0 0

interpulse delay

Figure 2.5: Electrical stimulation parameters.

Another property to be considered is the nature of the stimulation itself – namely, the contrast betweencurrent-regulatedandvoltage-regulated stimuli. Though some commercial electrical stimulators still apply the latter, it is not the most recomended. This is explained by the fact that what actually controls neural signals is electrical charge, which is more directly related with current than voltage. Since the impedance of the system being stimulated is not of resistive nature only, the current delivered will not be adequated. Also, if the impedance between the electrode and the neuron is too high, the charge delivered by voltage-regulated stimulators may be too weak to reach an stimulation threshold – thus, generating no response; by the same token, if the impedance is too small, this may incur in discomfort or even pain to the subject[5, p. 208]. A change of impedance during stimulation, thus, will result in undesired responses.

Followingly, as the signal being used is already settled for, one must decides on which sort ofmod- ulationwill be used. Modulation, in this situation, refers to how changes in an specific variable may be used to affect others. In this case, one must choose between modulating muscle force or sensations by changes in the waveform properties. Two main types of modulation are then presented:

pulse-width modulationandamplitude modulation. Despite that the differences in the sensations or contractions ellicited by them are usually very slight, the former is mainly implemented due to its

(29)

ease of execution, as generating high resolution in the time axis has generally a more straightforward implementation[44, p. 359].

In case of electrical stimulation on efferent pathways, there are specific matters that should be discussed more thoroughly. Knowingly, efferent pathways stimulation will have direct effect on muscle contraction, but how exactly a specific stimulus may affect the EMG response is a more complex matter. This is more easily explained byrecruitment curves, which graph the relation between a given stimulation parameter (e.g., pulse-width or amplitude) and amplitude of activation, which can be described by maximum EMG ellicited, generated force in newtons, highest peak-to-peak recorded value, etc. Of course, the relation shown can be represented in a three-axis plot, being two of them variables that contribute in stimulation and the third one, generally, once again related to the EMG value[45].

These curves are usually sigmoidal, being divided in three main sections: thedeadzone, where stimulation does not evoke any contraction; thehigh-slope, in which a behavior that can be approxi- mated to a linear function is found; andsaturation, the point at which any increase in stimulation will result in no higher activation in the muscle[46]. As noted before, it has been shown that an decrease in the interphase delay value will result in a greaterthreshold separation, meaning that the second section – the high-slope – will demonstrate a steeper slope. In parallel, an increase in this parameter provides a curve with better resolution and, thus, more fitting for applications involving different muscle activations[43]. Another significant usage of the recruitment curve is when the interface being used is that of an electrode array. In these cases, it is not known for sure how well each single channel of the array affects movement or sensation; in order to find out, one can make use of the recruitment curves for each of the electrodes and, based on them, define individualselectivity indexes, which should specify how well the targeted pathway is affected compared to the others[47, p. 443].

(30)

saturation zone

deadzone

high-slope

muscleactivation

parameter chosen

Figure 2.6: Expected effect of generated recruitment curve.

Still when considering recruitment curves, there is an interesting aspect that is worth mentioning related to the difference between physiological activation and artificial stimulation of muscles. Dur- ing natural activation, the nervous system recruits firstly type I small fibers, recruiting the larger fibers only if more force is needed. However, when dealing with external stimuli, the first motor units to be recruited will be the larger ones, since they have a smaller activation threshold. In terms of recruitment curves, this means that a natural recruitment curve is usually less steep than the one generated by electrical stimulation – which can come in the way of using applications that depend on different values[44, 48, p. 359, p. 159].

Another important aspect that should be considered when stimulating afferent pathways is one that may help deal with different muscle aspects: thecatchlike muscle property. This property is based on several changes in muscle tetanic contraction characteristics that follow a specific pattern of stimulation, described by an initial burst of two to four high-frequency pulses[49], which has been related to the increase in concentration of sarcoplasmic calcium ions, an important component of muscle contraction[50]. As mentioned, this property has been shown to affect different characteristics, one of which is muscle fatigue itself. It has been shown that it increases substantially resistance against decrease of EMG response after a prolongued time[3], which is mostly explained by the increase of calcium ions, which are closely related to muscle fatigue, as explained before. One more interesting

(31)

parameter that is affected by it is muscle potentiation, which are inversely related in matters of muscle strength; in other words, the usage of the catchlike property of the muscle in stimulation results in a smaller overall force during potentiation[21]. Other characteristics have also been related to catchlike muscle property – e.g. fiber type composition of the muscle[51]and muscle contraction[52]– proving its revelance to electrical stimulation.

A lot has been detailed in the electrical stimulation of efferent pathways; however, lots of it can be very useful insofar of explaining the stimulation in afferent pathways. For example, while efferent pathways stimulation is usually modulated by amplitude or pulse-width, one of the most common parameters used in neural encoding for sensations is frequency of stimulation [53, p. 362]Another one of these aspects is analogous to recruitment curves, the namedtuning curves. They are generated varying specific parameters of stimulation (usually the firing rate) and subjectively relating to the patient’s sensations, such as joint angle or pressure[54].

All these aspects that were deeply described will come in very handy when understanding how this work interfaced with subjects from different species and in different ways. A clear discernment of each characteristic of both encoding and decoding were fundamental in the development of the work and, hopefully, it was come accross appropriately.

2.3 Control of Arthropods Navigation

Once the basis for understanding neural interface with electrical stimulation of afferent pathways have been settled, studies that rely on these approaches for navigation control of arthropods will be discussed. These studies aim at controlling the direction at which insects and other invertebrates move, either by walking or even flying. In this section, we focus on research involving terrestrial insects, along with the methods proposed and results obtained.

Several studies have been made during the last decades considering neural stimulation of arthropods in order to control movement[55, 7, 56]. Even though our work aims at the implementantion of navigation control for insects, most of the works are focused on the development of electronic circuits light enough to be carried by the insect and which affect the overall movement the least possible. As this continuously shows to be a challenge, the navigation control used for arthropods has been mainly restricted to human remote controlling[1], though some researches on control algorithms have been implemented and will be detailed in subsection 2.3.2. In other words, most current researches only make use of navigation control when it comes to assessing the quality of electronic devices proposed.

However, a lot can still be learned about how stimulation aimed at motor control has been done.

(32)

remote operator

Figure 2.7: Example of remotely operated navigation control in a cockroach. Adapted from Latifet al [1].

2.3.1 Principles of Neurophysiology in Insects

Before entering in the navigation control realm, it is interesting to have a deeper understanding on how the nervous system of insects is organized, which is made available in[57]and shall be detailed in sequence.

Insects have simple nervous systems mainly composed by a dorsalbrainand segmentalganglia.

Ganglia are groups of interconnected neurons which, generally, represent similar functions. The brain is composed of three main structures: theprotocerebrum, thedeuterocerebrumand thetritocerebrum.

The first one is primarly associated with vision, as it is adjacent to the optic lobes, which is related to the orientation of antennas and biological cycle regulation. The second structure is associated with denser antennal information processing, as olfactory sensoring information. The last one, which surrounds the digestory system, is mainly associated with innervation of the labrum (upper lips).

Connected to the brain, there can be found a complex of fused ganglia calledsubesophageal gan- glion, which innervate different structures, such as mandibles, maxillae, and labium. In some insects, it also innervates hypopharynx, salivary glands, and neck muscles. The next structures are individual ganglia, which are connected – within eachsegment– bycommisuresand – among different segments

(33)

– by intersegmental connectives. Some of the ganglia connected arethoracic, which innervate both wings and legs for both afferent and efferent pathways. These are some of the structures primarly involved in locomotion. Theabdominal ganglia, by the same token, innervate abdominal muscles, which can be used in gait patterns focused on turning the insect around . Finally, thecaudal ganglion innervate the anus, genitalia and other sensorial receptors present incerci.

dorsal brain

protocerebrum deuterocerebrum tritocerebrum

subesophageal ganglia

thoracic ganglia

abdominal ganglia commissures

intersegmental connectives

Figure 2.8: Basic neurophysiology of insects.

In case of cockroaches, as the main insect used in our work, the main differences are the fusion of the subesophageal ganglion to the tritocerebrum and the specific nomenclature to the ganglia that innervate the legs: theprothoracicandmesothoracic ganglia.

2.3.2 Insects Artificial Navigation Using Neural Interfaces

Having understood each role within the nervous system in insects, the strategies used in the last decades are now detailed. Lots of them have been used throughout the years aiming at elliciting stimuli in invertebrates in order to control movement[1, 58]. Some of studies focus on stimulating the insect in specific regions of the Central Nervous System, which evoke coordinated motor activity[59]. These structure, when stimulated, chains a proper sequence of actions that allows the arthropod to execute functionalities such as walking[60]. All the while, several approaches can be followed in order to stimulate these systems, such as the usage of Command Neurons, specific neurons that trigger a whole behavior or part of a motor routine[60, 61]; stimulation of afferent strucutres, e.g. cockroaches hair plates, which set off specific patterns[58]. Another somewhat famous strategy, however, does not focus on the CPG: it consists on taking advantage of the mentioned evasive behavior of some arthropods (more specifically, cockroaches) and stimulate the antennas in order to provoke right and left turns[1, 2, 9].

One of these works placed a cockroach (wearing an electrical stimulation device) on top of an instrumented trackball[2], which allowed the researchers to compare kinematic data based on stimulation parameters. This resulted in valuable lessons about how stimulating the antenna can evoke specific motion and how consistent this motion can be. In fact, these researchers proposed in the end

(34)

of the work a line-following setup relying on two light sensors; unfortunatelly, due to great variance perceived in the motion, the results were not consistent.

turn (yaw)

forward(pitch)

Figure 2.9: Cockroach atop a trackball used for sensing. Adapted from Holzeret al[2].

Another work, however, showed to be very efficient in this matter. By using aMicrosoft Kinect sensor, the system could verify the cockroach’s position and orientation in time, allowing very fine control. The control implemented was focused on the angle difference between current orientation and desired orientation towards a waypoint. This way, the control allowed the cockroach to maintain a desired circular reference trajectory[9].

2.4 FES Control Using Evoked EMG

Another application involving the interface with the nervous system is control of movement. As muscle activation is intimately connected to movements, it is the main aspect that should be studied when implementing the control algorithm. In this sense, different variables for measuring the activation may be used, such as generated torque or joint angle; however, the most appropriate this work is evoked EMG. The usage of eEMG is justified by two main reasons: firstly, it provides a cleaner setup, in which sensors are attached in a more reliable way that interferes very little with movements.

Another good motive is the assessment of fatigue, that is done more precisely, since eEMG is a direct measurement of muscle activation.

Even though eEMG is the variable chosen in this work, implementations that rely on other variables have been tested, such as in[62], where aclosed-loop proportional and integral(PI) controller was implemented in order to control output force. This study was successful in dealing with muscle fatigue and potentiation effects by making use of force transducers as sensor inputs and modulating pulse-width in wire electrodes. By surgically dettaching the muscle from the body, the force transducer was connected to it and a wire EMG electrode was inserted into the muscle. The work made use of recruitment modulation, which directly mapped the relation between force and pulse-width.